Hunt security
Understand how hunts keep AI agents read-only and limited, handle attacker-controlled log content, keep findings visible, and what data goes to your AI provider.
Hunts let AI agents read your event data on a schedule, without a person watching. This page explains what the agents can and can't do, how nano handles log content that an attacker wrote, and what data leaves your installation.
Queries are read-only and limited
Every query a hunt runs on the server is read-only and limited:
- Only
SELECTqueries. A query must be oneSELECTstatement of up to 16 KiB, with no comments. nano refuses anything else before it runs. - The database enforces it too. Hunt queries run as a read-only database user with limits on time, rows, and memory for each query, and a required time filter.
- Small results. One query returns at most 200 rows to an AI model. Each run also has budgets.
A private, read-only connection
The server runner reads your data through its own private connection, a version of Agent Lane that only nano uses:
- Each runner gets its own short-lived database user, accepted only from the runner's own address. nano creates it when there is work and removes it when the runner is idle or stops.
- Hunt queries run at a lower priority than detections, search, and your analysts' own AI agents.
- nano records every query in the Agent Lane audit trail, labeled with the run, review, or campaign check that ran it.
Hunt queries read every log source except nano's own audit events. Source visibility doesn't limit them. What each analyst sees on the Hunts pages is still limited by that analyst's own source visibility.
AI agents can't change anything
The investigator, the reviewer, and the recheck agent can't write to nano or to your data:
| Rule | Why it matters |
|---|---|
| Agents can only report what they found. nano checks the report before saving it. | A model's claims are never saved unchecked. |
| nano reads every cited event again from the database. | A finding must be backed by real events. |
| Only the reviewer can escalate, and only with evidence it read. | A single model can't raise an alarm on its own. |
| Agents can't promote, dismiss, or close anything, or act on your systems. | A person makes every decision. In case mode the reviewer can open a case, but an analyst must acknowledge it. |
| Hunts never write detection rules. | A person ships rules through detection as code. |
| A campaign that an agent suggests starts dormant. | It does nothing until an analyst activates it. |
Attacker-written log content
An attacker can write text into your logs, including text that looks like instructions to an AI model. This is called prompt injection. nano treats all event data as data, never as instructions:
- Event data is marked as untrusted. nano wraps every result that contains event data in a marked block. It tells the model never to follow instructions inside those blocks.
- Only your team's text is trusted. The hunt's hypothesis and playbook, written by your analysts, are the only instructions the model follows.
- The reviewer decides first. It forms its verdict from the raw evidence before it can read the investigator's reasoning. Text that misled the investigator can't reach the reviewer the same way.
- Learned notes are labeled. Notes the investigator wrote are shown to later runs as untrusted, with where they came from.
- Notifications contain no AI text. An escalation notification carries a short reason that nano writes, never AI-written text or event content.
Findings can't be hidden by an AI agent
A manipulated model would most want to hide a real finding. nano prevents this:
- A finding that an investigator's suppression marks as a duplicate still appears on the Findings page, under Suppressed by the agent. Only an analyst's suppression hides a finding from the default view.
- An investigator's suppression lasts at most 30 days, and expires sooner if it stops matching.
- The investigator can only suppress a pattern it labeled benign in the same run.
- A reviewer verdict of benign counts only if it cites the finding's own events. Otherwise the finding stays a lead.
- AI agents never delete findings. Benign and duplicate findings stay on the Findings page with the reason for their label.
What data leaves your installation
Hunts send data only to the AI provider you choose. nano doesn't receive your data.
| Runner | What is sent | Where it goes |
|---|---|---|
| Server runner | The hypothesis and playbook, the list of sources, query results, what the hunt has learned, and the events the reviewer reads | The AI provider you set in Settings → AI & pivt, through your Cloudflare AI Gateway if you use one |
| Desktop runner | The same kinds of data, for the investigator | The analyst's AI agent and its provider |
To keep hunt data on your own network, use an on-premises AI endpoint that's compatible with OpenAI. To keep a Desktop hunt's evidence away from the server's AI provider, leave the server reviewer off for that hunt. It's off by default for Desktop hunts.
Audit events
| Event | Written when |
|---|---|
hunt.reviewer_escalated_case | The reviewer opened a case in case mode |
hunt_escalation_acknowledged | An analyst acknowledged an escalation |
hunt_escalation_mode_set | Someone changed a hunt's escalation mode |
hunt_ruling_recorded | An analyst recorded a ruling by hand |
hunt_lead_rereview_requested | An analyst asked for a new review |